基于混合分布时间序列的K-GRU建模

K-GRU Modeling Based on Mixed Distribution Time Series

混合分布的时间序列数据具有非平稳性、周期性等复杂的特性,且蕴含着未来的变化趋势,其复杂性对该类数据的预测精度造成了很大的影响。为了有效降低该类数据的预测误差,提出了一种K-均值与GRU神经网络相结合的混合模型——K-GRU混合模型,具有较高的预测精度。仿真实验表明,样本量为5000左右,K分别取2、3、4、5时,该模型均比GRU神经网络的预测效果好。模拟结果也与多项式模型、傅里叶序列和LSTM进行了比较,结果表明K-GRU混合模型的预测效果最好。最后将K-GRU混合模型应用于街边停车占用率数据中,进一步验证了该方法的有效性。

The time series data of mixed distribution has complex characteristics such as non-stationarity and periodicity,and contains the future trend.Its complexity has a great impact on the prediction accuracy of this kind of data.In order to effectively reduce the prediction error of this kind of data,in this paper,a hybrid model of K-means and Gru neural network k-GRU hybrid model is proposed,which has high prediction accuracy.The simulation results show that when the sample size is about 5000 and K is 2、3、4、5 respectively;the prediction effect of this model is better than that of GRU neural network.The simulation results are also compared with polynomial model,Fourier series and LSTM.The results show that the k-GRU hybrid model has the best prediction effect.Finally,the k-GRU hybrid model is applied to the street parking occupancy data to further verify the effectiveness of the method.